1. Introduction
Significant emphasis is currently placed on energy efficiency and its crucial role in accelerating the energy transition, in achieving the sustainable development goals (SDGs) promoted by the United Nations (
https://sdgs.un.org/, accessed on 30 November 2023), and in the implementation of the NetZero Roadmap [
1]. Although drastic measures must be taken at the national and supranational levels to achieve these goals, the advancement of technology has made it possible for individuals to actively contribute towards these goals, using tools accessible to all. For instance, the advent of smart devices, designed to interact with one another and with users, offers a wide range of possibilities for monitoring and regulating energy use, including in household settings. The increasingly widespread connection of home appliances, heating systems, lighting, and other devices to the smart grid enables not only remote control but also the collection of detailed consumption data. This increased accessibility of resources for better energy management can not only be financially convenient but also reveals a potentially crucial role in reducing environmental impact by promoting more sustainable behaviors. Nevertheless, research has shown that the adoption of such systems alone, without adequate support in reporting energy data and conveying the impact of such devices on consumption, does not generate significant results in terms of adoption of virtuous behaviors by the end users [
2,
3]. In this context, therefore, the crucial importance emerges of using effective feedback techniques, or eco-feedback (see
Section 2.1), to improve users’ understanding of the goals achieved, encouraging more responsible and efficiency oriented behaviors. Eco-feedback mechanisms give consumers a detailed view of their energy consumption patterns. The main objective is precisely to increase awareness about how their use of appliances and devices impacts energy expenses [
4]. Techniques based on natural language processing and generation, as well as conversational interfaces, fit into this context as potential eco-feedback processing tools that are meaningful to the user for efficient resource conservation. This potential is especially expressed in the possibility of integrating the provision of clear messages, and useful and personalized advice, with a level of human-like interaction achieved through dialogue, thus fostering greater acceptance of these tools as an aid to consumption efficiency.
Conversational agents have been used extensively in several related fields, with the aim of promoting sustainability and enhancing users’ commitment to adopt more eco-friendly habits; for instance, different examples of chatbots for the circular economy are reviewed in Zota et al. [
5], while other research projects have focused on the use of these tools to foster sustainable mobility beliefs [
6], to increase environmentally conscious behaviors among the employees of an office [
7], or to reduce food waste in a household [
8].
The goal of this survey is to provide an extensive overview of the state-of-the-art regarding conversational agents in the energy domain, with a focus on the use of tools that provide energy-related feedback through text or voice-based interactions. A large number of works can be found on visualization or gamification techniques devised for the delivery of eco-feedback (see
Section 2.1); likewise, countless articles have attempted to provide up-to-date reviews of conversational agents [
9,
10,
11,
12,
13], and have even applied them to specific sub-fields, such as health [
14], business [
15], and virtual reality and Internet of Things [
16], to name a few. To the best of our knowledge, however, the current article represents the first literature review on the use of conversational agents specifically to convey energy feedback. Our intent is to highlight the diversity of contributions and viewpoints involved in this domain, as well as to provide a clear organization of the methods and techniques found in the literature. This approach aims not only to describe what has been done so far, but also to serve as a starting point for future studies devoted to this topic, highlighting the potential and limitations of conversational agents in this context, as well as possible room for improvement and expansion. We thus aim to help create a solid foundation for future developments, pinpointing the challenges to be addressed and suggesting possible research directions to expand the understanding and effectiveness of such agents in the energy context.
In line with these goals and motivations, the present paper is organized so as to provide a preliminary background on the notions of energy feedback and conversational agents, also mentioning specific projects that lie at the intersection of these two domains (
Section 2). It will thus describe the protocol we adopted to conduct the survey (
Section 3) and will briefly outline the works identified through the search (
Section 4). A discussion (
Section 5) will then provide a more detailed account of the reviewed papers, while some closing remarks and suggestions (
Section 6) will complement the discussion, underscoring the current limitations of this type of research application and proposing potential future directions.
3. Survey Protocol
The general objective of this survey is to provide an overview of the studies in the literature that pertain to the use of conversational agents to provide energy feedback. In light of these premises, we thus intend to investigate four specific aspects, represented by the research questions formulated below:
RQ1: What is the primary focus of the works found in the literature dealing with conversational agents for energy feedback?
RQ2: What are the main goals (in terms of type of feedback and agent’s functionalities) and use scenarios for this type of conversational interface?
RQ3: What are the main approaches adopted for their development?
RQ4: In the existing body of work, what form of evaluation is carried out on conversational agents for energy feedback? What are the common practices (if any) and evaluation criteria?
Based on these research objectives, we conducted a comprehensive search of all the material available in the literature on the topic of conversational agents in the energy field, inspired by the recommendations established in the PRISMA 2020 statement for systematic reviews [
40]. We thus defined the pool of online sources to be consulted for the research of the material, we established inclusion and exclusion criteria that were in line with the general research objective of this review, and finally analyzed the collected material after an iterative selection process. Below we define this process in more detail, starting with the definition of the selection criteria and then describing the search procedure.
3.1. Inclusion and Exclusion Criteria
Our selection methodology focused on the identification of articles that investigate the topic of conversational agents developed with a particular emphasis on energy awareness and efficiency. Within this perimeter, we focused not only on the description of these agents but also on other related aspects, with the aim of understanding how such agents can be used to convey practical advice on energy savings and provide meaningful feedback on energy consumption.
User-centered design, accurate understanding of user’s messages and provision of appropriate responses, carefully considering the context of the conversation, or responding to requests using information from knowledge bases or external sources all represent only part of the complex dynamics involved in the development of a conversational agent and show how this topic lies at the intersection of multiple fields. This is especially true in the specific context of this study. Namely, due to the language-oriented distinguishing feature of this type of user interface, Natural Language Processing may play a central role in the development of such systems, but embedded and Internet-of-Things (IoT) devices serve as critical components as well (see
Section 2.2), in that they can provide real-time data, monitoring, and collecting information on energy use in a household or commercial/industrial building, even at the level of single appliances. In addition, contributions from the area of Human–Computer Interaction (HCI) can provide valuable insights. We have seen in
Section 2.1, for example, how the study of eco-feedback has been widely addressed in this field in order to identify effective mechanisms and strategies to communicate energy issues and highlight the main strengths and limitations of such feedback, as well as its actual long-term impact in terms of real changes in user habits. Likewise, user-centered design—through activities such as interviews, focus groups, prototyping, and usability testing—offers a diverse range of approaches to pinpoint potential areas of intervention, along with the needs, abilities, and values of the intended target users, and thus to improve the quality of interaction in terms of user expectations. In light of the above, the inclusion criteria adopted in this study were deliberately left open to research contributions from any research field, as long as they pertained to the use of conversational agents to communicate energy issues. Furthermore, and complementary to this, we also included some works identified through our systematic search that did not focus specifically on conversational agents as a whole, but rather on the development of systems aimed at handling at least one of the main tasks typically involved within a dialogue system (also summarized in
Figure 1). One such example is the generation of textual feedback from structured data, such as that from smart meters or other consumption monitoring tools. Although such systems have been developed as standalone contributions, they can potentially be integrated within a dialogue system and therefore represent a valuable and relevant contribution to this overview.
As regards the exclusion criteria, they were specifically conceived to ensure the consistency and specificity of our research focus. As a result, artifacts that rely primarily on visual and, more generally, non-language-based feedback have been excluded, as this deviates from our natural-language-oriented investigation. In line with this principle, we thus also decided to exclude IoT artifacts that do not involve user interaction through natural language, or the whole body of work that deals with the use of conversational agents for smart homes but do not have energy monitoring and feedback among their main goals or functionalities. As a matter of fact, although the availability of home automation systems can ultimately contribute to a greater energy efficiency, we believe that the use of chatbots or voice assistants for the mere automation of household appliances (i.e., offering functions for switching devices and appliances on or off) does not fully reflect the specific orientation of our research towards the promotion of energy awareness and efficiency through meaningful linguistic interactions. Finally, we did not consider as eligible contributions for the review the following categories: publications that have not been peer reviewed, such as pre-prints or Master’s theses, publications that were not fully accessible, and papers that do not introduce an original contribution (such as position papers). In addition, we limited our analysis to publications written in English, although some of them (as shown in
Section 5.3) may regard systems managing non-English conversations.
3.2. Information Sources and Search Protocol
As is customary in this type of research, some of the main databases and online search engines for scientific articles have been selected as primary data sources, namely ACM Digital Library, IEEE Xplore Science Direct, and Google Scholar. The research was carried out by querying these sources using a combination of the following keywords: (“chatbot” OR “conversational agent” OR “dialogue system” OR “natural language interface” OR “virtual assistant”) AND (“energy feedback” OR “energy consumption” OR “energy efficiency” OR “smart energy”). The search was carried out manually via the corresponding web interfaces, without resorting to APIs or automatized scripts. For none of the sources used for the search was a time constraint set on the date of publication, and in cases where the results provided by the query were of a large number (this was the case in particular for the ACM Digital Library and Google Scholar), only the results that could be consulted up to the fifth page (limit conventionally established after some preliminary tests) were considered for the selection, assuming that increasingly less relevant results would have been found after that limit. With this initial selection, we found a total of 508 records (250 in ACM, 22 in IEEE Xplore, 136 in Science Direct, and 100 in Google Scholar). The search was last updated in late October 2023.
Starting from the initial selection phase, we manually scanned this preliminary collection, only retaining for further inspection the articles we deemed potentially relevant to this survey. This filtering process generated an initial set of 47 items, subject to further assessment. The items were saved as records in a spreadsheet, with columns including the title, the source the paper was retrieved from, its URL, and whether it could be eligible for full screening on the basis of both the title and any keywords defined by the respective authors, along with an additional column with open notes explaining why a given paper should be excluded. During the analysis of this collection, cases of duplicates (i.e., the same article found in multiple databases) and articles that did not meet the peer review criteria or could not be considered as proper scientific contributions (e.g., dissertations) emerged. A preliminary evaluation of the abstracts of the remaining articles led to the exclusion of 12 of them, considered non-relevant as per the selection criteria outlined above, thus refining the final collection to 23 articles.
Finally, the last step involved a full screening of the articles in the selection and their final assessment with respect to the relevance criteria established for this study. At the end of this validation phase, 12 articles were considered fully relevant and included in the review. The spreadsheet created in the previous step was then extended to incorporate additional columns aimed at capturing specific aspects (discussed in
Section 5) that we intended to identify within the papers to address our research questions.
In order to ensure that potentially relevant papers were not overlooked in the aforementioned steps, additional research was conducted to assess possible contributions outside the main systematic search. Three contributions, already known to the authors of this manuscript, were included in the selection process. In addition, the works cited as related work in the 12 articles selected in the last phase of the systematic procedure were fully screened. Among these, excluding those already found in the systematic search, five articles were further analyzed. Out of these eight contributions found through non-systematic search, three were finally assessed as relevant and included in the review. This multi-source and iterative search protocol is also summarized in
Figure 2. In the next sections, we will discuss the results obtained after this process.
4. Overview of the Results
The collection used for this review consists of 15 contributions, 5 of which have been published in journals, and the remainder in academic conferences. These works have been published since 2015 but have been distributed more widely in 2018 and in the last three years, as also shown in
Figure 3. Below, we briefly outline the contributions we included in this review, also highlighting their primary strengths and weaknesses, while in the next section we will discuss more in-depth their main features in light of the research questions defined in
Section 3.
The authors develop a personal assistant that integrates into the conversational framework a multi-intent NLU component, a rule-based dialogue manager, and a planning module for the management of household appliances (mainly lights and HVACs) according to a Demand Response paradigm. The proof-of-concept was tested in a real-life scenario within the nZEB Smart Home project, carried out at the Center for Research & Technology, Hellas (CERTH) (
https://smarthome.iti.gr/, accessed on 30 November 2023); however, the feedback capabilities of the agent are quite limited and need to be extended.
Trivino and Sanchez-Valdes [
42]
The authors create a data-to-text generation system that generates textual advice on household consumption. The system is based on a model called GLMP (Granular Linguistic Model of Phenomena), for generating personalized text from numerical data. The model is designed to map quantities expressing energy consumption at different degrees of granularity into linguistic expressions. These converted expressions are then included in pre-defined text templates that provide custom advice on how to improve the daily energy consumption behavior in a specific household. The model introduced in this study uses actual consumption data from 12 households over a one-year period; however, it was at a very early stage at the time, and it needed additional experiments to assess its robustness.
Conde-Clemente et al. [
43]
This work builds upon the one by Trivino and Sanchez-Valdes [
42], and it expands it with the definition of the LDCP (Linguistic Descriptions for Complex Phenomena) framework. The generation model relies on a corpus of sentences that are useful for mapping linguistic expressions with the corresponding range of numerical values. The framework is applied on a real-world use case using consumption data obtained within a EU-funded project (NatConsumers, now ended (
https://cordis.europa.eu/project/id/657672, accessed on 6 December 2023); however, the use case presented in the paper is mainly for demonstration purposes. This suggests that the framework may not be fully mature for actual deployment in practical scenarios. Furthermore, as also acknowledged by the authors, the corpus of linguistic expressions may not be diverse enough, thus implying that some of the limitations of the system might have been mitigated with the addition of more linguistic data in the corpus.
The paper describes the creation of a virtual assistant, called GreenMoCa, that provides a fine-grained feedback on energy consumption and performs appliance switch operations upon user request. The system architecture consists of three main components: the conversational interface, an API service facilitating communication between the interface and the storage module responsible for saving the consumption data, and a separate API service enabling communication between the storage module and the connected smart plugs. The system underwent evaluation in a lab environment, with the participation of 10 individuals in the experiments. Nevertheless, it appears that it has not been tested in real-life scenarios.
In this work, the authors propose Cooee, a conversational interface to query consumption data over the energy management platform of La Trobe University, Australia. The back-end architecture of the interface comprises, besides the typical modules of a conversational agent, a pattern-matching-based module to retrieve a corresponding template of the given user’s request and a Text-to-SQL component that converts the identified template into a SQL query over the database of the energy platform. The interface thus displays the result of the query along with a textual explanation of the interpreted SQL query. From a preliminary evaluation, the Text-to-SQL module appears to offer promising results compared to other state-of-the-art models; however, for more effective dialogue management, the system could benefit from further integration with other natural language processing techniques.
This is a follow-up work from Gamage et al. (a) [
45], where Cooee’s functionalities are augmented with ChatGPT (
https://chat.openai.com, accessed on 30 November 2023) capabilities, precisely to address some of the main limitations encountered in the previous experiments. The addressed limitations concern, in particular, the use of a conversational context to tackle ambiguities, missing information in the user’s request, or for co-reference resolution. The system did not undergo a systematic evaluation, but an exploratory analysis of the chatbot’s outputs showed an adequate handling overall of these limitations. Additional tasks were also explored as integration into the Cooee system, benefiting from ChatGPT’s capabilities. Among these were the analysis of tabular data, the extraction of data from unstructured sources, and the generation of recommendations. This preliminary investigation revealed the importance of domain-specific knowledge in generating appropriate responses and avoiding the well-known hallucination problems common in generative models. Thus, the authors emphasize the need for appropriate prompt engineering strategies and domain-specific heuristics to ensure more effective integration of ChatGPT into their system.
This work does not describe a conversational agent in itself, but rather the development of a framework, named CANDY, to support the design of a sustainability oriented chatbot for the home environment. The framework definition resulted from the findings of a focus group comprised of experts from multiple domains (mainly computer science, energy sector, and psychology) and was aimed at eliciting, through an iterative discussion, the main dimensions to be considered in the design of a conversational agent, among which were the possible areas of intervention, the use cases, and the possible interaction methods, as well as engagement mechanisms. The generated framework could function as a valuable and practical foundation for developers. Nevertheless, it requires additional testing and validation.
The paper introduces Leafy, an app developed to encourage eco-friendly practices in households by integrating conversational agents into a smart-home setting. Specifically, the app integrates a multi-modal user–agent interaction—through a smart mirror—with gamification techniques to enhance the user engagement and their commitment to more sustainable behaviors in the domestic environment. However, the work, as described in the paper, lacks a proper evaluation of the system’s usability, and the authors explicitly recognize the need for additional exploration into the potential impact of conversational interactions on improving long-term engagement.
This paper describes the creation of a chatbot prototype aimed at providing real-time and personalized energy consumption information. The interface design is based on four principles inspired by Design Science Research [
50], and the prototype was built around two main scenarios specifically aimed at evaluating such principles. The evaluation results suggest that conversational agents are indeed a promising technology for providing energy feedback. However, the discussion with the experts highlighted the need to further explore the use of multi-modal, instead of text-only, agents. In parallel, the authors also recognize the importance of complementing the experts’ perspective with focus group sessions involving lay-users in order to gain a more diverse viewpoint on the design principles devised in this work. Finally, the study presented a simulated mock-up, but an actual implementation of such principles and the deployment of the developed agent might be beneficial for a thorough validation.
This research focuses on the study of the potential nudging effect achieved through the use of proactive virtual assistants for energy saving advice. The study involved simulations of interactions between users and a virtual assistant and showed that a significant number of users who initially provided neutral or mildly negative responses to the virtual assistant’s advice were more inclined to accept the advice after a follow-up interaction with the agent. Similarly to Gnewuch et al. [
49], the authors show the potentially positive impact of conversational interactions on promoting energy-saving behaviors. However, to further validate these findings a use case with a fully operational agent might be necessary.
The paper describes the development of a chatbot that provides information in English regarding the operation of the German energy system, starting from a model that schematically describes its functioning. The chatbot’s main end users are non-expert audience and policy-makers that need this type of data to make informed decisions. The conversational interface enables them to ask general questions about energy as well as more specific questions about the energy model it is connected to (a simplified model of the German energy system is used). Users can also alter the model with their input and compare the results obtained with both original and modified models. As mentioned above, the main strength of the agent consists of making a complex engineering model more accessible to non-technical users; on the other hand, the chatbot functionalities are closely tied to the characteristics and parameters of the particular energy system model it is designed for, and further adaptations would be necessary with a different energy model.
The authors describe a proof of concept designed to monitor the consumption and state of home appliances and send commands to switch them on and off. In addition to the typical modules of a dialogue system, an ontology is developed to describe the data concerning appliances’ consumption, the home environmental information, and the relation between the inhabitants and the appliances. A dedicated module is finally in charge of querying the ontology and retrieving the required information, building proper SPARQL queries. The system design also emphasizes aspects related to accessibility issues and restriction of actions, or if they are deemed inappropriate for certain users, such as young children. However, extensive evaluations from users are necessary to assess such design principles, particularly in relation to the accessibility issues.
The authors introduce a web application that includes both a visual (a wall-mounted tablet) and a voice (using Alexa) interface that provide feedback aimed at helping users make better use of the thermostat, while at the same time promoting community-wide energy conservation through gamification mechanisms. The feedback mechanism is based on mathematical models that define (1) an energy conservation behavior score that quantifies users’ efforts in reducing energy use and (2) an algorithm for personalized action recommendations along with their potential impacts. In an extensive user study, participants reported the positive influence of the personalized feedback in improving their thermostat settings. On the other hand, throughout the experiments the user engagement with the voice assistant was reportedly less frequent compared to the visual interface, thus revealing potential limitations in the conversational agent’s design.
Santos Fialho et al. [
54]
The paper introduces PowerShare VA, a virtual assistant aimed at informing users about their energy usage in different forms, both with a voice-based application and through a web interface. The tool has been designed to provide consumption feedback on multiple temporal time frames (i.e., with a daily, weekly, and monthly dashboard). Real-world experiments with users in different scenarios demonstrated the assistant’s capabilities in effectively interacting with users and providing meaningful feedback. On the other hand, only one participant per scenario was involved in the experiments, thus affecting the generalizability of the findings. Furthermore, the final discussion with participants elicited the need to build more user-tailored functionalities.
The authors create a chatbot that communicates the household energy consumption (e.g., concerning general trends or consumption peaks). An additional mail-alerting module is also integrated into the system to inform the user whenever consumption exceeds a given threshold. The chatbot uses consumption data obtained from an openly accessible dataset. The dataset comprises energy consumption readings for a sample of 5567 London households participating in the UK Power Networks-led Low Carbon London project between November 2011 and February 2014. However, although the paper mentions a comparison of chatbot performance, there is a lack of informative details about how this comparison was actually conducted.
7. Conclusions
In this article, we provided an overview of the literature regarding conversational agents and feedback generation in the energy sector. We primarily remarked that there is a vast amount of research on providing feedback through visualizations or using gamification techniques to influence user behaviors. In contrast, work that focuses on user engagement through conversational tools and interfaces is more limited. We attribute this to the very specific nature of these conversational agents on the one hand, but also to the selection criteria we decided to follow on the other. Concerning the latter in particular, we adopted very strict criteria, precisely in order to focus on work that encompassed both of the two key points that were in line with our research focus, i.e., the delivery of eco-feedback through dialogue-based interactions; this greatly narrowed the range of eligible works. Although we cannot exclude that the various filtering and screening steps were influenced by unintentional biases, we believe that, given the specialized nature of our research topic, the selected number of articles is still representative and it adequately captures the diversity of this type of conversational agent.
In addition to the methodological considerations, there are several implications of our findings for this field that should be considered, as outlined in
Section 5.5. First, we have highlighted the importance of taking an inter-disciplinary approach in the development of conversational agents for eco-feedback. This translates, for example, into a greater centrality given to the user in both the design and evaluation phases of these agents. With regard to the potential capabilities of the agents at hand, a further implication could involve greater synergy than that evidenced in the reviewed papers between eco-feedback and automation capabilities. This aspect aims to expand the agents’ operational capabilities when it comes to a more efficient management of energy resources and devices. Finally, the diversity of usage scenarios, currently oriented predominantly towards domestic applications, suggests that developers might explore and design conversational agents with features suitable for broader contexts, including commercial and industrial settings. In parallel, we deem it critical that future developers consider the possibility of a shift towards multimodal settings and, more generally, towards making these tools more accessible to a wider audience (e.g., elderly or visually-impaired people). The significant advances in machine learning and NLP, combined with the growing adoption of smart devices on one side and the greater acceptance of conversational agents in everyday tasks on the other, makes the potential of these tools for energy purposes even greater.
In conclusion, despite the relatively limited contribution given to conversational agents in the energy context to date, we believe that their role will become increasingly central in supporting users in the efficient management of energy consumption, thus helping to promote greater awareness and sustainability in the use of energy resources, both individually and collectively.